Clustering With EMand KMeans
Neil Alldrin
Department of Computer Science
University of California,San Diego
La Jolla,CA 92037
nalldrin@cs.ucsd.edu
Andrew Smith
Department of Computer Science
University of California,San Diego
La Jolla,CA 92037
atsmith@cs.ucsd.edu
Doug Turnbull
Department of Computer Science
University of California,San Diego
La Jolla,CA 92037
dturnbul@cs.ucsd.edu
Abstract
Two standard algorithms for data clustering are expectation maximiza
tion (EM) and Kmeans.We run these algorithms on various data sets
to evaluate how well they work.For high dimensional data we use ran
dom projection and principal components analysis (PCA) to reduce the
dimensionality.
1 Introduction
The KMeans algorithm ﬁnds
clusters by choosing
data points at random as initial
cluster centers.Each data point is then assigned to the cluster with center that is closest to
that point.Each cluster center is then replaced by the mean of all the data points that have
been assigned to that cluster.This process is iterated until no data point is reassigned to a
different cluster.
EM ﬁnds clusters by determining a mixture of Gaussians that ﬁt a given data set.Each
Gaussian has an associated mean and covariance matrix.However,since we use shperical
Gaussians,a variance scalar is used in place of the covariance matrix.The prior probability
for each Gaussian is the fraction of points in the cluster deﬁned by that Gaussian.These
parameters can be initialized by randomly selecting means of the Gaussians,or by using
the output of Kmeans for initial centers.The algorithm converges on a locally optimal
solution by iteratively updating values for means and variances.
2 Low Dimensional Data Clustering
For the ﬁrst part of our project,we implemented the EM and KMeans algorithms.Our
implementations were tested on two sets of twodimensional data:a distribution generated
by two Gaussians and an annulusshaped distribution.
2.1 KMeans on TwoDimensional,Two Gaussian Data
The KMeans algorithmworks very well on this data set,effectively converging in three or
four iterations (see ﬁgure 1).
2
0
2
2
1
0
1
2
3
Iteration 1
2
0
2
2
1
0
1
2
3
Iteration 2
2
0
2
2
1
0
1
2
3
Iteration 4
Figure 1:The progress of the KMeans algorithmwith
and randominitialization on
the twoGaussian data set (note:some data points omitted for clarity).
2.2 EMon TwoDimensional,Two Gaussian Data
The EMalgorithmalso performs well,typically converging within 5 iterations (see ﬁgure
2).
5
0
5
4
2
0
2
4
Iteration 1
5
0
5
4
2
0
2
4
Iteration 4
5
0
5
4
2
0
2
4
Iteration 7
Figure 2:The progress of the EMalgorithm with
and random initialization on the
twoGaussian data set (note:some data points omitted for clarity).The radius of the circle
around each Gaussian is set to its variance.
2.3 KMeans on TwoDimensional,Annulus Data
On the annulus data,KMeans also works well,with the centers convergingto points evenly
distributed around the annulus in four or ﬁve iterations (see ﬁgure 3).
2.4 EMon Two Dimensional,Annulus Data
On the annulus data set,the EMalgorithmalso performs well,converging within 10 itera
tions (see ﬁgure 4).
1
0
1
1
0
1
Iteration 1
1
0
1
1
0
1
Iteration 2
1
0
1
1
0
1
Iteration 3
1
0
1
1
0
1
Iteration 4
1
0
1
1
0
1
Iteration 5
1
0
1
1
0
1
Iteration 6
Figure 3:The progress of the KMeans algorithmwith
and randominitialization on
the annulus data set (note:some data points omitted for clarity).
We verify that our code for EMis progressively ﬁnding a better ﬁt for the data by checking
that the negative log likelihood after each iteration never increases.As can be seen in Figure
5,this value decreases after each iteration.
3 High Dimensional Data Clustering
Most realworld data sets are very highdimensional.However,the performance of cluster
ing algorithms tends to scale poorly as the dimension of the data grows.For this reason the
dimensionality of data sets is often reduced by various techniques before it is clustered.
Our data set is very highdimensional,since each data point is a 240 x 292 image with
256 shades of gray.Treating each pixel as a dimension yields a 70080dimensional data
set,which makes clustering difﬁcult given our computing resources.To reduce the dimen
sionality of our data set,we experimented with random projections and principal compo
nent analysis (PCA).Randomprojections have the desirable property that highly eccentric,
highdimensional Gaussians become more spherical when projected down to a small ran
dombasis.
Our data set is a collecton of images of the faces of 14 different people expressing different
emotions.Each person was instructed to make a happy,sad,surprised,afraid,disgusted,
and angry face.Our primary goal is to classify the facial expression of a given image by
clustering our data set into six clusters,one for each emotion,and then calculating which
cluster is most likely to contain that image.We are also interested in clustering our data set
2
0
2
2
1
0
1
2
Iteration 1
2
0
2
2
1
0
1
2
Iteration 2
2
0
2
2
1
0
1
2
Iteration 4
2
0
2
2
1
0
1
2
Iteration 6
2
0
2
2
1
0
1
2
Iteration 7
2
0
2
2
1
0
1
2
Iteration 9
Figure 4:The progress of the EMalgorithm with
and random initialization on the
annulus data set (note:some data points omitted for clarity).The radius of the circle around
each Gaussian is set to its variance.
0
2
4
6
8
10
12
14
16
18
20
3000
3200
3400
3600
3800
4000
4200
Iteration
Error Function
Figure 5:The negative log likelihood of the EM algorithm on the annulus data
set with
and random initialization.This error function is
.
into 14 clusters,one for each person,and conducting the analogous experiment of classi
fying images by person.We were concerned that clustering images to distinguish between
emotions would ﬁnd clusters of different people,rather than different facial expressions.
To avoid this,we make use of an image of each person making a “neutral” face.We add
a”difference image,” deﬁned as the difference between each image of a person expressing
an emotion and that person’s neutral face to the data set.This set is used to cluster based
on facial expressions,whereas we use the raw images to classify particular people.Our
intuition was that clustering by person would be more successful than clustering by facial
expression.
We downsample all images by a factor of 64 (8 in the x and y dimensions) to reduce
the effects of noise.Intuitively,the downsampling does not remove information crucial
to clustering since a human can still identify people and their facial expressions at this
resolution.
3.1 Classifying People
Supervised Clustering
0
20
40
60
80
100
120
0
2
4
6
8
10
12
14
image number (lines seperate images of same person)
cluster number
Figure 6:Supervised clustering,1080dimensional data (no dimension reduction).
Note about ﬁgures 6 and 7:The graphs show how each image was classiﬁed,by cluster
number on the vertical axis,and the images on the horizontal axis.These images are
sorted by known category,where each category is seperated by vertical dashdot lines.
A square indicates a point classiﬁed by ﬁnding the clustercenter (mean) to which that
point is closest.A dot indicates a point classiﬁed by ﬁnding the Gaussian with the highest
probability at that point.
As an initial test,we use all the images of a particular person to calculate a maximum
likelihood mean and variance of that person,and then use these 14 Gaussians to classify
each image in the data set.We use two classiﬁcation methods,one classiﬁes a data point
by ﬁnding the mean to which that point is closest,the other ﬁnds the Gaussian with the
highest probability at that point.This supervised clustering test represents an upper bound
on howwell we can expect unsupervised clustering algorithms to perform.
Figure 6 is the result of ﬁtting 14 Gaussians to the raw image data and then trying to
classify each image.Evaluating Gaussians on such highdimensional data causes numerical
precision problems,so we are unable to use the probability method to classify images.
0
20
40
60
80
100
120
0
2
4
6
8
10
12
14
image number (lines seperate images of same person)
cluster number
Figure 7:Supervised clustering,15dimensional data (randomprojection).
Notice that only three (of 110) images are misclassiﬁed,indicating that the raw data are
well seperated.
Figure 7 is the same as ﬁgure 6,except the data were pojected down to a random 15
dimensional basis.These lowdimensional data are still well seperated,but there are a
few more misclassiﬁcations than in the highdimensional data.Reducing the number of
dimensions with PCA yields comperable results.
Unsupervised Clustering
0
20
40
60
80
100
120
0
2
4
6
8
10
12
14
image number (lines seperate images of same person)
cluster number
Figure 8:Unsupervised clustering,1080dimensional data (no dimension reduction).
0
20
40
60
80
100
120
0
2
4
6
8
10
12
14
image number (lines seperate images of same person)
cluster number
Figure 9:Unsupervised clustering,15dimensional data (randomprojection).
0
20
40
60
80
100
120
0
2
4
6
8
10
12
14
image number (lines seperate images of same person)
cluster number
Figure 10:Unsupervised clustering,15dimensional data (PCA projection).
Note about ﬁgures 8 through 10:The squares indicate the KMeans results and the dots
indicate the EMresults.Some EMresults are not present due to numerical precision prob
lems.
Figure 8 is the result of running KMeans (EMfailed due to numerical precision problems)
on the entire highdimensional data set,looking for 14 clusters,and classifying all the data
points according to the clusters those algorithms found.About ten clusters were found by
KMeans that correlate well to distinct people.
Figure 9 is the same as ﬁgure 8 except the data have been projected down to a 20 dimen
sional random basis and EMresults are included.About ﬁve clusters were found by both
algorithms that correlated well with distinct people.
Figure 10 shows the clusters found by KMeans and EMon the data set projected down to
15 principal components.EMclusters marginally better than KMeans in this case.It also
appears that the PCA basis worked better than the randombasis.
3.2 Classifying Facial Expressions
Supervised Clustering
0
10
20
30
40
50
60
70
80
90
100
0
1
2
3
4
5
6
image number (lines seperate images of same expression)
cluster number
Figure 11:Supervised clustering,1080dimensional data (no dimension reduction).
0
10
20
30
40
50
60
70
80
90
100
0
1
2
3
4
5
6
image number (lines seperate images of same expression)
cluster number
Figure 12:Supervised clustering,15dimensional data (randomprojection).
To see if disjoint clusters of facial expressions exist in our data set,we use all the difference
images of each facial expression to calculate a maximumlikelihood mean and variance of
that expression,and then use these six Gaussians to classify the images.We use the same
two classiﬁcation methods as in section 3.1,classifying a data point by ﬁnding the mean to
which that point is closest and the other ﬁnding the Gaussian with the highest probability
at that point.This supervised clustering test represents an upper bound on howwell we can
expect unsupervised clustering algorithms to perform.
Figure 11 shows that most data points can be identiﬁed with the correct cluster,indicating
the raw data can be partitioned into distinct clusters of facial expressions.
Figure 12 is the same as ﬁgure 11 except the data have been projected down to a 15
dimensional random basis.About a quarter of the data are misclassiﬁed,indicating the
clusters in the projected data are less distinct.
Unsupervised Clustering
0
10
20
30
40
50
60
70
80
90
100
0
1
2
3
4
5
6
image number (lines seperate images of same expression)
cluster number
Figure 13:Unsupervised clustering,1080dimensional data (no dimension reduction).
Figure 13 is the result of runningKMeans (EMfailed due to numerical precision problems)
on the entire highdimensional data set,looking for 6 clusters,and classifying all the data
points according to the clusters found.The only cluster that correlates well with a particular
facial expression is cluster 1 corresponding to the happy expressions.
Figure 14 is the same as ﬁgure 13 except the data have been projected down to a 20 dimen
sional random basis and EM results are included.Again,the only cluster that correlates
well with a particular facial expression is the cluster associated with happy expressions,
but this cluster is not as disjoint (from other clusters) as in the highdimensional data set,
since there are more false positives and misses.
Figure 15 shows the clusters found by KMeans and EM on the data set projected down
to 20 principal components.As in the previous two ﬁgures,only happiness correlates well
with a particular cluster.There are slightly fewer misses than with the high dimensional
data;however,there are many more false positives.
0
10
20
30
40
50
60
70
80
90
100
0
1
2
3
4
5
6
image number (lines seperate images of same expression)
cluster number
Figure 14:Unsupervised clustering,20dimensional data (randomprojection).
0
10
20
30
40
50
60
70
80
90
100
0
1
2
3
4
5
6
image number (lines seperate images of same expression)
cluster number
Figure 15:Unsupervised clustering,20dimensional data (PCA projection).
4 Conclusion
Downsampling the images greatly improved clustering in all cases.We suspect this is
because most noise was averaged out.
We tried two techniques to reduce the dimensionality of our data set,projecting it down to a
lowdimensional randombasis,and principal component analysis.Both of these techniques
degraded the quality of the clustering,but made the expectation maximization algorithm
feasible.Our observation is that PCA was only marginally better,if at all,than a random
projection despite its computational intensity.
In general,we noticed that KMeans performs comparably to EM;however,EM fails on
highdimensional data sets due to numerical precision problems.Another problem with
EMis that Gaussians often collapsed to delta functions.Our technique to prevent this was
to reset the variance of collapsed Gaussians to a more reasonable value,and to set the mean
of those Gaussians to randomdata points.
We would have liked to run our clustering algorithms on our data sets and then validate
the results by classifying novel data,however when we reserved a portion of our data for
validation,the clusters the algorithms found did not correspond at all to the classes we were
trying to ﬁnd.We strongly suspect this is due to the lack of enough sample points to deﬁne
accurate Gaussians.
5 Future Work
The technique we use to convert images to feature vectors is simply to list all of the pixels in
the image.Clearly,this is the a naive approach because it ignores the correlations between
neighboring pixels.Our ﬁrst attempt was to create feature vectors that were the Gabor
wavelet transforms of each image using 4 scales and 6 rotations.This took a long time
and produced a feature vector that was far too large for our computational reasources.
Future work should explore the potential of this approach by using a more efﬁcient wavelet
transformprocedure.
In addition to KMeans and EM,KHarmonicMeans is another clustering algorithm that
could be used to classify images.For each cluster center,KHarmonicMeans computes the
harmoic mean of the distance to every data point,and update that cluster center accordingly.
This algorithmis less sensitive to initial cluster centers than KMeans,but does not have the
problemof collapsing Gaussians exhibited by EM.For these reasons,KHarmonicMeans
might ﬁnd better clusters in highdimensional data.
Individual Contributions
Neil Alldrin:
My primary contribution was writing the EMcode (most of the em
* ﬁles).This included
a lot of effort devoted to how to handle collapsing Gaussians and how to prevent divide by
zeros caused by lack of numerical precision (which was only a problem on high dimen
sional data).I also generated the graphs for the 2dimensional data and helped write the
latex document you are nowreading.I had lots of fun.
AndrewSmith:
I learned Matlab.I wrote the script to load the images (loadFaces.m).I wrote the function to
downsample images.I wrote Kmeans.I wrote the scripts to generate the highdimensional
data graphs (section 3) in this paper.I wrote a function to generate a Gabor ﬁlter bank,
using code from[2] to evaluate Gabor functions.I experimented with using Gabor wavelet
transforms for feature vectors of images,but it was too slow.I had lots of fun.
Doug Turnbull:
My primary focus in this assignment was to design and implement various experiments for
high dimensional data using kmeans and EM.These tests included implementing random
projection and PCAprecompuation algorithms,creating scripts to run tests for various data
sets,and collecting results for analysis.Developing these tests were often a nontrival task
due to a large number of parameters (projection matrices,clustering algorithm,data sets,
etc.) that greatly affect the quality of the results.I had a little bit of fun.
References
[1] Bishop,C.M.(1995) Neural Networks for Pattern Recognition,Oxford University Press,New
York.
[2] Manjunath,B.S.&Ma,W.Y.(1996) Texture Features for Browsing and Retrieval of Image Data,
IEEE PAMI,vol 18,no.8,pp.837842.
[3] Sanjoy,D.(1999) Learning Mixtures of Gaussians,IEEESymposiumon Foundations of Computer
Science (FOCS).
[4] Zhang,B.(2000) Generalized KHarmonic Means,HewlettPackard Laboratoris Technical Re
port.
Enter the password to open this PDF file:
File name:

File size:

Title:

Author:

Subject:

Keywords:

Creation Date:

Modification Date:

Creator:

PDF Producer:

PDF Version:

Page Count:

Preparing document for printing…
0%
Σχόλια 0
Συνδεθείτε για να κοινοποιήσετε σχόλιο